Selection of Dataset for Emotion Detection with Respect to Federated

Learning

G. K. Jakir Hussain

and G. Manoj

Division of Electronics and Communication Engineering, Karunya Institute of Technology and Sciences, Karunya Nagar,

Coimbatore 641114, Tamil Nadu, India and Department of Electronics and Communication Engineering, KPR Institute of

Engineering and Technology, Coimbatore 641407, Tamil Nadu, India

Division of Electronics and Communication Engineering, Karunya Institute of Technology and Sciences, Karunya Nagar,

Coimbatore 641114, Tamil Nadu, India

Keywords: Emotion Detection, Federated Learning, Dataset, Data Diversity, Human Computer Interactions.

Abstract: Emotion detection (ED) plays a vital role in applications like health, human–computer interaction (HCI), and

personalization. Federated learning (FL) has the potential to provide robust models for ED while keeping data

private across distributed clients. This work has considered the critical factors that influence dataset selection

for FL on ED tasks. The range of dataset types, class balance for varying emotional states, and attention to

privacy considerations that safeguard users' sensitive information are among the important factors to consider.

Therefore, several datasets are examined in terms of how well they extent the range of emotional expressions

and replicate actual client data distributions. This study highlights datasets such as FER-2013 for photos,

RAVDESS for audio, and ISEAR for textual data are highly relevant for the construction of generalized ED

models in FL setups that simulate near-real-world settings, as per review publications. The FL technique can

balance the data diversity and privacy preservation and hence saves a pathway to harness collective

intelligence from distributed data sources while ensuring ethical practices for handling data. Finally, FL is

becoming a critical topic for addressing issues in identifying ED by selecting the suitable datasets.

1 INTRODUCTION

Emotion detection (ED) is a process that identifies

and classifies human emotions into modalities such as

text, audio, and images. This requires methods for

text by way of natural language processing (NLP)

methods and those for audio through speech analysis,

facial expression recognition (FER) from images, and

more advanced ones that use deep learning (DL) and

machine learning (ML) algorithms. Specifically, in

ED through federated learning (FL), models are

trained across decentralized devices, ensuring user

privacy since data remains local while model updates

are aggregated at a central location. It helps in

improving mental health support, enhancing human

computer interaction, and making the marketing

strategies more personalized. Facilitating early

identification of emotional disorders, it makes

Artificial Intelligence (AI) empathetic and

personalizes marketing campaigns according to the

user's emotions for better engagement and customer

satisfaction.

FL enables the training of decentralized models

across a number of devices while maintaining local

privacy of data, and at the same time, it centrally

aggregates updates to models for better performance.

FL is an approach to ML whereby model training

takes place across decentralized devices holding local

data samples, without transferring data to a central

server (G. K. J. Hussain and G. Manoj, 2022). Key

principles include data privacy, where data resides on

local devices; iterative model updates, where only

model parameters or gradients are shared and

aggregated to create a global model; and security

features enhancing scalability. The interaction

between ED and FL gives privacy-preserving models

that can analyze emotional states from decentralized

devices, hence providing personalized experience to

users with the safety of sensitive data. FL applied to

ED faces several obstacles, including heterogeneity in

device data, ineffective communication, and bias

against the model due to unequal data distributions.

Selecting a dataset for ED in FL involves privacy,

diversity, and representativeness. The dataset needs

Hussain, G. K. J. and Manoj, G.

Selection of Dataset for Emotion Detection with Respect to Federated Learning.

DOI: 10.5220/0013917400004919

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Research and Development in Information, Communication, and Computing Technologies (ICRDICCT‘25 2025) - Volume 4, pages

593-599

ISBN: 978-989-758-777-1

593

text, speech, and facial data for emotional cues.

Anonymized data with consent is crucial for privacy.

Diverse and representative datasets aid model

performance across demographics. Public datasets

like IEMOCAP and Sentiment140 may be adapted for

FL. Data partitioning is essential for non-independent

identity (non-iid) data in FL to ensure effective

generalization.

2 EMOTION DETECTION

ED involves the identification and subsequent

classification of human feel through various inputs,

either text, speech, or facial expression. The

techniques from NLP, Computer Vision, and ML are

attached in analysis of emotional signals. ED can be

categorized into three main types. ED can express

feelings like joy or anger through text, speech, and

facial recognition. Text-based detection uses NLP for

sentiment analysis, while speech-based methods

analyze tone and pitch in audio. Facial expression-

based detection employs computer vision to identify

emotions through facial movements. These

approaches can be used combined or separately to

enhance accuracy of ED systems in applications like

customer service and mental health monitoring.

2.1 Types of Datasets

Type, diversity, and balance of emotional label are

important considerations when choosing datasets for

FL based on ED. For a strong model training, one

would desire datasets that can be expressive about

various journeys for emotional expression. Privacy is

of the utmost importance, and datasets must adhere to

regulations controlling the suppression of sensitive

data. A dataset that can enable realistic client data

distributions is also necessary in order to simulate

real-world scenarios in FL. FER-2013 is used for

images, RAVDESS for audio, and ISEAR for textual

data. All of these offers decentralised learning and are

helpful in acquiring some understanding of the

emotional state.

2.1.1 Text Based ED

Selecting a dataset for text-based ED should be done

in accordance with its relevance, balance, and

diversity. It will involve a combination of labeled and

emotive content, for example, for social media posts

or reviews, or primed conversations. These range

from widely used and richly emotional-annotation

ones like ISEAR (International Survey on Emotion

Antecedents and Reactions) and EmoReact. The real-

world scenarios should be represented in the dataset

for the model to build resilience. Moreover, datasets

would be split to simulate scenarios of real-world FL

enabling decentralized training of a model preserving

confidential and private information associated with

several users.

2.1.2 ISEAR Dataset

The ISEAR dataset is a unique database of 7,666

emotional statements contributed by 1,096

participants belonging to different cultural

backgrounds. The dataset can be used in creating an

Emotion Dominant Meaning Tree for the

classification of emotional statements to obtain

higher precision with increased emotion categories.

The algorithms will be trained to learn how to identify

and classify expressed emotions using the ISEAR

Dataset from statements provided by participants.

About 60% of the dataset is used in constructing the

Emotion Dominant Meaning Tree. Jain and Asawa

(Jain, S. and Asawa, K, 2019). mention the following

strengths of the ISEAR dataset: it holds a large

number of reports from respondents who expressed a

wide range of emotions within a variety of settings. A

deep analysis of emotional states is possible to be held

due to the links between each emotion and certain

event assessments and responses. It subsequently

assures the model's effectiveness by selecting

relevant records to increase the elicitation conditions'

accuracy, thus proving to be a valuable dataset in case

scenarios of real-world tests and validation of

elicitation rules.

Table 1: Emotions described in ISEAR dataset. (Source:

author).

Types of Emotions Number of examples

Anger 1096

Disgust 1096

Fea

1095

Sadness 1096

Shame 1096

Joy 1094

Guil

1093

Total 7666

This is based on the ISEAR dataset, which

Alotaibi (Alotaibi, F.M., 2019) used to recognize

emotions from text with logistic regression, which

has since gone on to be categorically distinguished

between fear, joy, sadness, guilt, and humiliation.

Assessment metrics include the F1-score, precision,

and recall. Better performance can be seen with

logistic regression compared to the rest: support

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

594

vector classifier, K-nearest neighbor, and extreme

gradient boosting. Table 1 displays the emotions

described in ISEAR dataset.

Asghar et al. (2019): The primary aim is the

internet information identification of emotions

through the effective utilization of supervised ML

algorithms. It considers, in an assessment of the ML

classifier, ISEAR data that contains 5477 reviews,

basically categorized as joy, fear, sadness,

humiliation, and guilt; this is after applying various

preprocessing techniques on them, such as

tokenization and stop word removal, before applying

the classifier. Table 2 shows the detail description of

ISEAR dataset and figure 1 depicts the types of anger

based on the graphical representation of emotions

from ISEAR dataset.

Table 2: Detail description of the ISEAR dataset. (Source: author).

Citation Variet

lanation Real-world Pertinenc

[2]

Various settings for eliciting

emotions are modeled

Assumed to be high because

of the construction of the

nitive-emotive

Relevant to cognitive

systems that simulate

emotions

[3]

Moderate, encompassing

fundamental feelings

Manual annotation with

distinct emotional descriptors

Helpful for classifying

emotions using logistic

regression

[4]

High, based on many web

sources

High, utilising methods for

supervised ML

Appropriate for identifying

feelings in a variety of web

material

[5]

Various conditions for eliciting

emotions are modelle

High, making use of BERT's

titudes

Efficient at recognising

emotions from text

Figure 1: Graphical representation of the emotions in

ISEAR dataset.

Among the many uses for the dataset in Adoma et al.

2020 is the assessment of machine learning models

for emotion categorization. This dataset is ISEAR,

consisting of 7666 sentences labeled for seven

emotions. By making the dataset balanced, training

and testing are possible without mitigation for class

imbalance. As such, this dataset will have several

preprocessing tasks such as tokenization or stop word

removal. Results from this study may be compared to

results from other studies in the same field,

particularly serving as a benchmark for future ED

studies with the application of ML models.

2.1.3 Speech Based ED

The audio-visual data in IEMOCAP dataset consists

of 12 hours, which contains data from 10 actors in 5

sessions. It contains data that was both improvised

and scripted, and it is divided into four emotional

classes: neutral, happy, sad, and angry. Three to four

assessors comment each utterance, and they award

labels by majority vote. This multimodal dataset

makes use of text, audio, and motion capture (Mocap)

information. Its vast nature and good quality make it

extremely important to research on ED. Using the

IEMOCAP dataset, research by Tripathi and Beigi

focuses on multi-modal emotion identification. This

dataset allows neural network models for robust

emotion recognition to be trained and evaluated. It

consists of a variety of audio, video, and text data

from dyadic interactions. The 1440 audio-only WAV

files in the RAVDESS Emotional Speech Audio

dataset have a sampling rate of 16 bits @ 48 kHz.

Recorded in a neutral North American accent, it

contains 24 professional actors 12 female, 12 male

delivering lexically-matched phrases. Emotions

manifest at two intensity levels (normal and strong),

with a neutral expression in between. Calm, pleased,

sad, furious, afraid, surprised, and disgusting are

among the emotions. The RAVDESS dataset has

been widely used in recent research to improve voice

emotion recognition techniques. Table 3 displays the

Comparison of the four datasets based on Speech

based ED type.

Selection of Dataset for Emotion Detection with Respect to Federated Learning

595

Table 3: Comparison of the Four Datasets Based on Speech Based Ed (Source: Author).

Dataset IEMOCAP RAVDESS CREMA-D EmoDB

Explanation Emotion speech database

Emotional speech

audio

Emotional speech Emotional speech

Source

USC Institute for

Creative Technologies.

Ryerson Audio-

Visual Database of

Emotional Speech

Crowd-sourced

Technische

Universität Berlin

Contributors 10 actors 24 actors 91 actors 10 actors

Reactions

Anger, Happiness,

Sadness, Neutral,

Surprise, Frustration,

Excitedness, and Fear.

Calm, Happy, Sad,

Angry, Neutral,

Fearful, Disgust,

Surprised.

Angry, Disgusted,

Fear, Happy,

Neutral, Sad,

Surprise

Anger, Boredom,

Disgust,

Anxiety/Fear,

Happiness, Sadness

A varied dataset called CREMA-D has 7,442 original

audio clips with 91 performers representing a range

of demographics (48 men and 43 women, ages 20 to

74, and different races). Twelve standardized words

representing six different emotions Anger, Disgust,

Fear, Happy, Neutral, Sad across four intensity ranges

Low, Medium, High, and Unspecified were uttered

by each actor. This dataset has a broad demographic

representation and is used for voice emotion

recognition research.

An innumerable of research into emotion

recognition depends on the CREMA-D dataset.

Shahzad et al. utilize CREMA-D for multi-modal

deep learning, where text and audio inputs are fused

to capture complex emotional expressions. The

variation of this dataset substantially enhances model

generalization across different modalities and thereby

improves both the accuracy and the system's

robustness in recognizing emotions.

The EMODB database developed by the Institute

of Communication Science at the Technical

University of Berlin which is an open-source German

emotional database. It contains 535 utterances from

ten professional speakers five of whom are male and

the other five are female. The seven emotions covered

by this database are: neutral, disgust, happiness,

sorrow, anxiety, anger, and boredom. The data was

down-sampled to 16 kHz for standardization after

being initially captured at a 48 kHz sampling rate.

2.1.4 Facial Recognition-Based ED

Another The process of facial recognition-based ED

trains algorithms to recognize human emotions from

facial expressions using datasets like the Extended

Cohn-Kanade Dataset (CK+) , the Facial Expression

Recognition 2013 (FER2013), the Multimodal

Database of Emotionally Salient Stimuli (MMI), the

Japanese Female Facial Expression (JAFFE) Dataset,

and the Affect Net Dataset. Labeled datasets for

posed and naturalistic expressions are provided by

CK+ and FER2013. MMI offers audio-visual signals

for analyzing emotions, and JAFFE offers insights

through the expressions of Japanese women.

The CK+ dataset includes 920 photos that have been

modified and taken from the original CK+ dataset.

These pictures have been pre-processed using the

`haarcascade_frontalface_default’ method to provide

standard 48x48 pixel grayscale dimensions and face

cropping. A Haar classifier was used to filter noisy

images that were impacted by things like skin tone

fluctuations, hair artifacts, and room illumination in

order to improve clarity and identification. The

dataset is organized into three columns for each entry:

pixel values (2304 per image), emotion label (which

has predefined indices for various emotions), and

usage classification into three categories: training

(80%), public test (10%), and private test (10%).

Table 4 shows the Demonstration of the CK+ dataset

and the number of samples.

Table 4: Demonstration of the CK+ dataset and the number

of samples (Source: Author).

Types of Emotions in CK+

dataset

Number of

Samples

Ange

Dis

ust 59

Fea

iness 69

Sadness 28

Surprise 83

Neutral 593

Contem

t 18

The FER2013 dataset, which includes approximately

35,000 grayscale photos classified into seven

emotional categories anger, disgust, fear, pleasure,

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

596

sorrow, surprise, and neutral is used for FER. It was

chosen specifically for use in FER task training and

evaluation models. Every image of 48x48 pixels is

labeled with different emotion categories.

Researchers frequently utilize FER2013 to

benchmark algorithms for applications involving

computer vision and emotion identification. A

number of DL-based facial ED studies are thus

heavily reliant on the FER-2013 dataset. By

incorporating audio, visual, and textual modalities,

the Multimodal Emotion Lines Dataset (MELD)

builds upon the Emotion Lines dataset. It features

numerous speakers and more than 1,400 lines and

13,000 utterances taken from the Friends television

series. Every word is annotated both with sentiment

positive, negative, or neutral along with an emotion:

Anger, Disgust, Sadness, Joy, Neutral, Surprise, or

Fear. The dataset includes 213 TIFF-formatted, high-

resolution grayscale pictures of 10 Japanese female

expressers wearing 7 posed facial expressions,

including neutral. Every expression has different

photos from each expresser, for a total of many

images per emotion. It contains 60 Japanese viewers

averaged semantic judgments of six different face

expressions. Table 5 shows the Comparison of the

five datasets based on the Facial recognition-based

ED.

Table 5: Comparison of the five datasets based on the facial recognition-based ED (Source: Author).

Dataset CK+ [10] FER2013 MMI JAFFE AffectNet

Source Kaggle Kaggle Kaggle Kaggle Kaggle

Determination FER FER ED FER FER

Quantity of

Images

593 35,887

18,000 (video

frames)

213 1,000,000+

Image Category Grayscale, posed

Grayscale,

posed

Video frames

(multimodal)

Grayscale,

posed

RGB, diverse

Resolution Various 48x48 pixels Various 256x256 pixels Various

Reaction types

8 (fear, anger,

disgust, surprise,

contempt,

happiness, neutral,

sadness)

7 (fear, angry,

surprise,

disgust, happy,

neutral, sad)

7 (anger, fear,

surprise,

disgust, sadness,

neutral,

happiness)

7 (disgust,

anger, sadness,

happiness,

neutral, surprise,

fear)

8 (fear,

neutral,

surprise,

happy, anger,

sad, contempt,

disgust)

Notation FACS action units

Crowd-sourced

labels

Video-based

labels

Human labeled

Crowd-

sourced labels

2.1.5 Highly Used Datasets in ED

A wide range of datasets specialized to various

modalities, including text, speech, and video, are

extremely beneficial to emotion recognition research.

Every dataset has a unique function: SemEval-2018

task 1 employs labelled tweets for text-based emotion

analysis, whereas IEMOCAP offers emotional

annotations in voice scenarios. For tasks involving

the recognition of emotions in videos, Emo React is

perfect because it comes with annotated clips that

cover a wide range of emotions. MELD facilitates

study into multimodal emotion comprehension by

integrating emotion labels with textual dialogues

from the Friends TV series. Emo Bank provides

dimensional emotion rankings that are taken from

blog sentences and social media, which gives a

distinctive viewpoint. The versatility of SEMAINE's

records, which allow for the examination of

emotional states in text, video, and audio modes, is its

main strength. These varied datasets are essential for

creating and assessing ED models, which advances

NLP and affective computing. Federated learning is

the most common approach for emotion detection

because it maintains decentralized data, improves

model robustness by utilizing a variety of data

sources, and facilitates collaboration among

dispersed clients without jeo pardising sensitive data.

The figure 2 shows some challenges in ED Dataset.

Selection of Dataset for Emotion Detection with Respect to Federated Learning

597

Figure 2: Challenges in Ed Dataset.

2.2 Federated Learning

FL is a decentralized ML technique that uses local

input and allows several devices or edge servers to

cooperatively train a common model. Local

computation and aggregates are used to compute

model changes rather than transferring data to a

central server. This lowers communication costs

while protecting user privacy. FL comes extremely

handy in situations like healthcare, IoT, and mobile

applications when there are several, diverse datasets

dispersed over various places. It solves problems like

connectivity problems and data privacy laws, making

it appropriate for applications needing customised

models without sacrificing the confidentiality of

individual data.

2.2.1 Federated Learning: Dataset

Necessities

In FL, datasets must meet requirements for effective

and secure decentralized model training by ensuring

privacy and security through encryption and secure

aggregation techniques. Differential privacy adds

noise to data or model updates to protect against re-

identification. Following FL protocols like Federated

Averaging (FedAvg) shares only model parameters,

not raw data, during updates. Handling data

heterogeneity in FL involves addressing non-

independent and identically distributed data with

personalized models, data clustering, or federated

transfer learning. Robust aggregation techniques like

Trimmed Mean, Median or Krum help manage

outliers or skewed data distributions. To enhance

model performance, adaptive learning rates make

adjustments in response to data heterogeneity.

Protocols adapting to network conditions reduce

communication overhead and transmission

frequency. Scalability in FL supports a large number

of devices by distributing workload, using

hierarchical FL with local coordinators, and

employing efficient resource management and load

balancing techniques. Addressing these requirements

enables effective leveraging of decentralized datasets

while maintaining privacy, handling data

heterogeneity, ensuring communication efficiency,

and scaling to support numerous devices.

2.2.2 Analysis of the Federated Learning ED

Dataset

FL analyses datasets dispersed across several devices

without centrally aggregating them, particularly in

emotion recognition tasks. This method allows for

cooperative model training while continuing data

privacy. Assuring data consistency across

heterogeneous sources, controlling device

heterogeneity, and addressing potential biases

generated by different data distributions are some of

the key issues in analyzing such datasets. These

difficulties are lessened by FL, which permits local

model updates on every device. These changes are

then combined to improve a global model without

requiring the raw data to leave the devices.

Preprocessing stages include data normalization,

feature extraction and maybe data augmentation to

improve model generalization in order to analyze

such datasets successfully. Recurrent neural networks

(RNNs) for sequential data, such as text or time-series

physiological signals, or CNNs for image-based

emotions are examples of ML models appropriate for

FL in ED. To sum up, in order to analyze ED datasets

for FL, it is necessary to undertake thorough

preprocessing, choose suitable models, and pay close

attention to privacy and performance measures that

are specific to distributed learning settings.

2.2.3 Problems in FL Dataset Selection

Making sure datasets abide with privacy laws like the

GDPR and getting consent from data owners to use

and share their information safely in FL. In FL,

striking a balance between a variety of representative,

heterogeneous data sources and preserving data

quality and schema uniformity throughout all

involved clients. Collaborative learning: identifying

and reducing biases to guarantee fair model

performance across many demographic groups.

Managing the communication overhead associated

with FL as well as restrictions on the processing,

memory, and storage capabilities of client devices. FL

ICRDICCT‘25 2025 - INTERNATIONAL CONFERENCE ON RESEARCH AND DEVELOPMENT IN INFORMATION,

COMMUNICATION, AND COMPUTING TECHNOLOGIES

598

frequently involves handling identically dispersed

and non-independent data in an efficient manner to

provide reliable model training. Ensuring sure

datasets may be split and distributed among numerous

clients in an efficient manner while preserving

successful training.

3 CONCLUSIONS

Choosing an appropriate dataset for ED is essential to

developing dependable, widely applicable models

that preserve anonymity. Datasets like ISEAR,

IEMOCAP, RAVDESS, CREMA-D, EmoDB, CK+,

FER-2013, MMI, JAFFE, AffectNet, SemEval-2018,

and SEMAINE provide a great variety of modalities:

text, audio, and facial expressions. These datasets

form the basis for an in-depth study of emotion

recognition. For example, more insightful audio

samples into very emotional speech come forth from

IEMOCAP and RAVDESS, whereas the ISEAR

dataset is rich with textual data in terms of emotional

responses. Again, some of these datasets are focused

on facial expressions relevant to visual ED, such as

FER-2013, CK+, and JAFFE. These datasets can be

used by several clients due to the fact that FL is

innately distributed, which will not risk user privacy

and promote data diversity in order to increase model

performance. Comprehensively including multi-

modal datasets ensures an all-round approach in

emotion identification to improve the accuracy of

capturing the complexity of human emotions.

However, class disparities should be dealt with and

the databases should represent real situations. FL can

significantly advance the science of ED by thoughtful

selection and class balancing of these datasets, so that

more morally correct applications be derived in health

and other fields that are more personalized. Future

applications of emotion detection with FL include

promising enrichments of human-machine

interactions and protections of the privacy of users.

This develops in terms of multimodal integration,

privacy, scalability, cultural diversity, and real-time

applications.

REFERENCES

Adoma, A.F., Henry, N.M., Chen, W. and Andre, N.R.,

December. Recognizing emotions from texts using a

bert-based approach. In 2020 IEEE 17th International

Computer Conference on Wavelet Active Media

Technology and Information Processing (ICCWAMTI

P), (2020) 62-66.

AffectNet Dataset: https://www.kaggle.com/datasets/ngot

hienphu/affectnet

Alotaibi, F.M., Classifying text- based emotions using logi

stic regression. VAWKUM Transactions on Computer

Sciences, 7(1), (2019) 31-37.

Asghar, M.Z., Subhan, F., Imran, M., Kundi, F.M.,

Shamshirband, S., Mosavi, A., Csiba, P. and Varkonyi-

Koczy, A.R., Performance evaluation of supervised

machine learning techniques for efficient detection of

emotions from online content. arXiv preprint

arXiv:1908.01587 (2019).

CK+ Dataset: https://www.kaggle.com/datasets/davilsena/

ckdataset

CREMA- D Dataset: https://www.kaggle.com/datasets/ejl

ok1/cremad

EmoDB Dataset: https://www.kaggle.com/datasets/piyush

agni5/berlin-database-of-emotional-speech-emodb

FER2013 Dataset: https://www.kaggle.com/datasets/msa

mbare/fer2013/code

G. K. J. Hussain and G. Manoj, Federated Learning: A

Survey of a New Approach to Machine Learning, 2022

First International Conference on Electrical, Electronic

s, Information and Communication Technologies (ICE

EICT), Trichy, India, (2022) pp. 1-8.

IEMOCAP Dataset: https://www.kaggle.com/datasets/sa

muelsamsudinng/iemocap-emotion-speech-database

JAFFE Dataset: https://www.kaggle.com/code/mpwolke/j

apanese-female-facial-expression-tiff-images

Jain, S. and Asawa, K., Modeling of emotion elicitation

conditions for a cognitive- emotive architecture. Cogn

itive Systems Research, 55, (2019) 60-76.

MMI Dataset: https://kaggle.com/datasets/zaber666/meld-

dataset

RAVDESS Dataset: https://www.kaggle.com/datasets/uw

rfkaggler/ravdess-emotional-speech-audio

Selection of Dataset for Emotion Detection with Respect to Federated Learning

599